Machine Learning-Driven Insight into Crystallization Kinetics in Liquid Crystals

Understanding the crystallization behavior of materials is crucial for controlling their structural and functional properties, particularly in systems exhibiting mesophases such as liquid crystals. In this work, we present an integrated approach that combines polarized light microscopy (PLM) with deep learning to quantitatively analyze the crystallization dynamics of the liquid crystalline compound 9BA4.

We trained a convolutional neural network based on the U-Net architecture to perform semantic segmentation of PLM images, allowing for automatic identification and distinction between crystalline (Cr) and smectic C (SmC) phases observed during non-isothermal cooling. The model generates pixel-wise probability maps for each phase, which are subsequently binarized to compute the degree of crystallization as a function of temperature.

To characterize the crystallization kinetics, a sigmoidal function was fitted to the experimental crystallinity–temperature curves. The inflection point of the fitted function was used to determine the temperature of maximum crystallization rate. This automated workflow significantly reduces subjectivity and manual effort compared to traditional texture analysis methods.

Our results highlight the potential of combining classical optical microscopy techniques with modern deep learning tools to extract quantitative, reproducible insights from complex phase transition phenomena. The proposed methodology can be adapted to other systems showing texture evolution during thermal processing, paving the way toward high-throughput and objective studies of crystallization in soft matter and beyond.